Dynamic credit limit management method and system based on multi-dimensional user credit portraits
By assessing user risk through multi-dimensional credit profiling and deep neural network models, and dynamically adjusting credit limits, the shortcomings of credit limit management are addressed, enabling accurate assessment and flexible adjustment, reducing risk, and improving user satisfaction and resource utilization efficiency.
Patent Information
- Application Number
- CN202510824942.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-19
- Publication Date
- 2025-10-28
AI Technical Summary
Existing credit limit management methods lack comprehensiveness and accuracy in credit assessment and cannot be dynamically adjusted. This results in users with good credit not receiving reasonable credit lines, while users with high credit risk receive loans beyond their repayment capacity. Furthermore, fixed credit limit strategies increase the risk for financial institutions.
By collecting data through multi-dimensional user credit profiles, a credit profile is constructed. Combined with a deep neural network model, the user's risk level and volatility coefficient are assessed. The pre-approved credit limit and interest rate are dynamically matched. The credit limit is adjusted using multi-objective optimization, and the model parameters are optimized through closed-loop feedback.
It enables accurate assessment of users' creditworthiness, reduces credit risk, improves resource utilization efficiency, adapts to market changes, provides flexible credit services, and enhances user satisfaction and the competitiveness of financial institutions.
Smart Images
Figure CN120852033A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of credit assessment and credit management technology, specifically to a method and system for dynamic credit limit management based on multi-dimensional user credit profiles. Background Technology
[0002] As one of the core businesses of financial institutions, credit business faces tremendous opportunities and challenges. With the rise of internet finance, the user base is constantly expanding and becoming more diversified. Traditional credit approval models can no longer meet market demands. Credit assessment, as a key link in credit business, directly affects the risk control and profitability of financial institutions. Accurately and comprehensively assessing users' credit status and reasonably determining credit limits are of great significance for reducing credit risk and improving the efficiency of financial resource allocation.
[0003] Currently, existing technologies have many problems in user credit assessment and credit limit management. On the one hand, most credit assessment methods rely only on limited data such as users' historical credit records and income statements, which are difficult to comprehensively reflect users' true credit status and repayment ability. This single-dimensional data collection and analysis method cannot effectively identify potential credit risks and is prone to misjudgment. As a result, users with good credit may not receive reasonable credit limits, while users with high credit risk may receive loans beyond their repayment capacity. On the other hand, in terms of credit limit management, existing systems often adopt fixed credit granting strategies and lack dynamic adjustment mechanisms. They cannot flexibly adjust in a timely manner according to changes in users' credit status and fluctuations in the market environment. When users' economic situation, consumption behavior, or macroeconomic conditions change, fixed credit limits may expose financial institutions to greater risks and also limit the service and development of high-quality customers.
[0004] In conclusion, existing credit limit management methods have significant shortcomings in terms of the comprehensiveness and accuracy of credit assessment and the dynamism of credit limit management. There is an urgent need for a new technological solution to address these issues, adapt to the needs of financial technology development, and enhance the competitiveness and service quality of financial institutions. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a dynamic credit limit management method and system based on multi-dimensional user credit profiles. It can connect with data from multiple channels to collect multi-dimensional user data, and then perform in-depth analysis and feature extraction on this data to construct a comprehensive and detailed user credit profile. Compared to traditional assessment methods that rely on only a few dimensions of data, it can more accurately assess a user's credit status and repayment ability. Based on this, a risk assessment model is used to calculate credit risk data by combining multiple factors, providing a scientific basis for determining pre-approved credit limits and pre-transaction interest rates. This multi-dimensional and precise credit assessment system effectively reduces credit risk, avoids bad debts and losses caused by inaccurate assessments, and also allows financial institutions to allocate credit resources more rationally and improve resource utilization efficiency.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a dynamic credit limit management method based on multi-dimensional user credit profiles, the specific steps of which are as follows:
[0007] S100, Multi-dimensional User Credit Data Collection and Fusion: User credit feature data is acquired in real time through a distributed data interface, and feature engineering is used to clean and normalize the user credit feature data to generate a vectorized feature set for the user credit profile.
[0008] S200, Risk-linked credit scoring modeling: Input the feature set of S100 into the pre-trained risk assessment model, and output the user's risk level and risk volatility coefficient;
[0009] S300, dynamic matching of pre-approved credit limit and interest rate: Based on the risk level of S200, the credit rule engine is called to generate an initial credit limit range, and the dynamic interest rate fluctuation range is calculated in combination with the risk volatility coefficient;
[0010] S400, Credit Limit Fitting and Dynamic Adjustment Generation: Multi-objective optimization is performed on the initial credit limit range and interest rate range of S300 to fit the dynamic credit limit, and the real-time market interest rate factor is introduced to generate the final dynamic adjustment amount.
[0011] S500, closed-loop feedback and model iteration: The actual credit performance data is fed back to the risk assessment model of S200 and the optimization algorithm of S400. The model parameters are updated through an online learning mechanism, forming a closed-loop link of credit assessment - credit decision - dynamic adjustment - feedback optimization.
[0012] Furthermore, the S100 user credit feature data includes static data, dynamic data, and scenario data, wherein:
[0013] The static data includes: the user's age, gender, education, occupation, proof of income, proof of assets, and credit records;
[0014] The dynamic data includes: real-time consumption behavior data, social network activity, mobile device usage behavior, and credit rating of social network contacts;
[0015] The scenario data includes: the unemployment rate of the user's industry, the GDP growth rate of the macroeconomy, and the amount of regional credit adjustment.
[0016] Furthermore, the risk assessment model is jointly trained using historical data and time-series user behavior data, dynamically adjusting the weight allocation of different dimensions to output the user's risk level and risk volatility coefficient. The specific training process is as follows:
[0017] Training data preparation: Collect historical data and user behavior time series data. Divide the historical data into time windows. The data in each time window includes the user's historical static data and scene data within that time period. The user behavior time series data includes the user's historical dynamic data within that time period. Preprocess the historical data and divide it into training set, validation set and test set in a ratio of 7:2:1.
[0018] Risk assessment model construction: A deep neural network is used as the risk assessment model. This model consists of an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the vectorized feature set generated by S100, and it is used to receive user credit feature data. The hidden layers contain multiple neurons, and a non-linear transformation of the input data is achieved through an activation function, f(x) = max(0,x), where x is the input of the neuron. This function effectively alleviates the gradient vanishing problem and improves model training efficiency. The number of neurons in the output layer is set according to the number of risk level categories; that is, the number of neurons in the output layer is the same as the number of risk level categories and is set to K. The output layer uses a softmax function for normalization. Where z is the input vector of the output layer neuron, K is the number of output layer neurons, and σ(z) j This represents the probability of the j-th neuron's output. This indicates that the input z of the j-th neuron in the output layer is... j Perform exponential operations with the natural constant e as the base. It applies to all K neurons in the output layer. The values are summed, and the probability distribution of users belonging to different risk levels is obtained through the softmax function. In the neural network, the output 'a' of each neuron is calculated from the input x, weight W, and bias b. For the l-th neuron, its z... l =W l a l-1 +b l a l =f(z)l ), where z l It is the weighted input of this layer, a l-1 It is the output of the previous layer, W l and b l These are the weight matrix and bias vector of the layer, respectively, and f is the activation function;
[0019] Model training process: The cross-entropy loss function is used as the optimization objective. The cross-entropy loss function is... Where N is the number of training samples, K is the number of risk level categories, and y qj This represents the true label of the q-th sample belonging to the j-th risk level. Let represent the probability that the model predicts the q-th sample belongs to the j-th risk level. The model is trained using stochastic gradient descent. In each training iteration, samples are randomly selected from the training set, and the gradient of the loss function for each sample is calculated with respect to the model parameters weights W and bias b. The gradient is calculated as follows: The model parameters are updated based on the gradient. Where α is the learning rate, which controls the step size of parameter updates. By continuously adjusting the model parameters and evaluating the model performance on the validation set, a pre-trained risk assessment model is obtained.
[0020] Furthermore, the risk level predicted by the model The specific calculation process is as follows: The feature set of the training samples is input into the input layer of the deep neural network. The input layer passes the data to the first hidden layer, and the hidden layer calculates the result based on z. l =W l a l-1 +b l and a l =f(z) l The input data is weighted, summed, and processed by an activation function, and then passed layer by layer until it reaches the output layer. At the output layer, the input vector z is obtained through weighted summation and then processed by the softmax function. The probability of the sample belonging to each risk level is calculated, and the vector composed of these probabilities is the model's prediction output, where the value of the j-th element is...
[0021] The feature set generated by S100 is input into the pre-trained risk assessment model. The output layer obtains the probability distribution of the user belonging to different risk levels through the softmax function, selects the risk level with the highest probability as the user's risk level, and maps this probability value to the preset risk range of five levels from A to E.
[0022] For the calculation of user risk volatility coefficient, based on the characteristics of the input user behavior time series data, the standard deviation in time series analysis is used to measure the degree of volatility of user behavior. That is, the user's behavioral characteristic sequence within this time period is x1, x2, ..., x n The mean of this feature sequence is Then the risk volatility coefficient of this feature This serves as the user's ultimate risk volatility coefficient.
[0023] Furthermore, in S300, the pre-approved credit rule engine is invoked to generate an initial credit limit range. The specific process is as follows:
[0024] Risk Level to Credit Limit Range Mapping: The risk levels obtained from S200 are mapped to the credit limit ranges in the pre-approved credit rule engine. Based on historical data, the pre-approved credit rule engine sets a corresponding initial lower limit L for each risk level. min and upper limit L max Specifically:
[0025] For users with a risk level of A, the lower limit of the initial credit limit range Up to 80% of the user's historical annual income. 120% of the user's historical annual revenue;
[0026] For users with a risk level of B, the lower limit 60% of the user's historical annual income, with a maximum limit. 100% of the user's historical annual revenue;
[0027] For users with a risk level of C, the lower limit 40% of the user's historical annual income, with a maximum of This represents 80% of the user's historical annual revenue.
[0028] For users with risk level D, the lower limit 20% of the user's historical annual income, with a maximum of 60% of the user's historical annual revenue;
[0029] For users with a risk level of E, the lower limit 10% of the user's historical annual income, with a maximum of This is 30% of the user's historical annual revenue.
[0030] Furthermore, the S300 dynamic interest rate fluctuation range is determined based on the risk volatility coefficient and the market benchmark interest rate. When the market benchmark interest rate is r0 and the risk volatility coefficient is σ, the interest rate adjustment coefficient is k for each risk level. The higher the risk level, the larger the value of k. For users with risk level A, k... A =1.0, for users with risk level B, k B=1.2, for users with risk level C, k C =1.4, for users with risk level D, k D =1.6, for users with risk level E, k E =1.8;
[0031] For the dynamic interest rate lower bound r min = r0 + k × σ, dynamic interest rate ceiling r max =r0 + 2k × σ, thus obtaining the dynamic interest rate fluctuation range (r min , r max );
[0032] At the same time, when the risk volatility coefficient σ is greater than the preset threshold σ th The system triggers a supplementary material application instruction, requiring users to submit proof of income, asset collateral documents, and third-party guarantee information. During the supplementary material application stage, the initial credit limit range is frozen and a 72-hour countdown begins. If the user fails to submit the materials within the time limit, the credit granting process is automatically closed. When the materials are approved, the risk score of S200 is updated based on the new data, and S300 is re-executed.
[0033] Furthermore, S400 performs multi-objective optimization by constructing a multi-objective optimization function. This function sets a risk coverage objective R and a profit maximization objective P. The risk coverage objective R measures the degree to which the credit line covers risk. Where w u It is the weight of the u-th risk level, g u This represents the risk probability corresponding to the u-th risk level, where U is the number of risk levels. The profit maximization objective P is used to comprehensively consider factors such as loan interest income and funding costs. Among them l v This is the limit for the vth loan product, r v It is the interest rate of the vth loan product and r v ∈[r min ,r max ], c v Let be the cost of funds for the v-th loan product, and V be the number of loan products. Then the multi-objective optimization function is F = αR + βP, where α and β are weight coefficients used to balance the importance of each objective, and α + β = 1.
[0034] The multi-objective optimization function F is solved, and during the solution process, the initial quota interval [L] generated in S300 is used. min ,L max and interest rate range [r] min ,r maxBy adjusting the combination of credit limit and interest rate, the multi-objective optimization function F is maximized to determine the optimal dynamic credit limit L. op ;
[0035] A real-time market interest rate factor and a user behavior feedback factor are introduced. The real-time market interest rate factor M reflects the real-time changes in market interest rates and is obtained by collecting interest rate data of similar financial products in the market and performing a weighted average calculation. The user behavior feedback factor B integrates the user's actual behavior after obtaining the pre-approved credit limit; that is, B = 1 when the user repays on time and 0 otherwise. Based on the user's behavior data, the user's credit status is assessed in real time, and the assessment result is transformed into the user behavior feedback factor B. The final dynamic adjustment amount ΔL = L op ×(M+B).
[0036] On the other hand, the dynamic credit limit management system based on multi-dimensional user credit profiles consists of the following components: data collection and fusion module, risk assessment module, pre-credit decision module, dynamic fitting module, and closed-loop feedback module.
[0037] The data acquisition and fusion module deploys a distributed data crawler and API gateway, integrates internal and external data sources, and has a built-in feature engineering engine to output a vectorized feature set of user credit profiles.
[0038] The risk assessment module includes an offline training unit and an online prediction unit. The offline unit trains the risk assessment model using historical data, while the online unit calculates the user's risk score and risk volatility coefficient in real time and generates risk-driven decision signals.
[0039] The pre-credit decision module has a built-in credit rule engine that sets an initial credit limit range based on decision signals, supports manual intervention interfaces and automated material supplementation request generation, and outputs a credit limit and interest rate range pair with constraints.
[0040] The dynamic fitting module uses multi-objective optimization to perform secondary calibration on the pre-credit results, and combines real-time market data and user behavior data to output dynamic adjustment amounts and credit activation instructions for manual review.
[0041] The closed-loop feedback module collects user behavior data after credit authorization, constructs a feedback dataset, triggers a model retraining task, and synchronizes the updated model parameters to the risk assessment module and the dynamic fitting module.
[0042] Compared with existing technologies, this dynamic credit limit management method and system based on multi-dimensional user credit profiles has the following advantages:
[0043] I. This invention collects multi-dimensional user data and conducts in-depth analysis and feature extraction to construct a comprehensive and detailed user credit profile. This profile enables a more accurate assessment of a user's creditworthiness and repayment ability. Based on this, a risk assessment model is used to calculate credit risk data in conjunction with multiple factors, providing a scientific basis for determining the pre-approved credit limit and pre-transaction interest rate. This multi-dimensional and precise credit assessment system effectively reduces credit risk and avoids bad debts and losses caused by inaccurate assessments. At the same time, in response to fluctuations in market interest rates and changes in the macroeconomic situation, the credit limit can be adjusted in real time. This dynamic management approach ensures that the credit limit matches the user's actual creditworthiness and the market environment.
[0044] Second, this invention fully considers various factors such as changes in users' real-time credit data, fluctuations in market interest rates, and changes in the macroeconomic environment. When users' consumption behavior and credit data of repayment records change, the system can promptly capture and calculate the dynamic adjustment amount of the credit limit. This not only ensures that the financial institution's risk is controllable under different market conditions, but also provides users with more flexible and reasonable credit services, thereby improving user satisfaction and loyalty.
[0045] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0047] Figure 1 A flowchart illustrating the operation of a dynamic credit limit management method based on multi-dimensional user credit profiles;
[0048] Figure 2 This is a diagram showing the module composition of a dynamic credit limit management system based on multi-dimensional user credit profiles. Detailed Implementation
[0049] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0050] Example 1
[0051] This embodiment details the operation of a dynamic credit limit management method based on multi-dimensional user credit profiles. By collecting multi-dimensional user data, a credit profile is constructed. A risk assessment model is used to determine the risk level and volatility coefficient, and then the pre-approved credit limit and interest rate are matched. Through multi-objective optimization, a dynamic credit limit is fitted and dynamically adjusted. At the same time, the model is continuously optimized with the help of closed-loop feedback, which effectively improves the accuracy and flexibility of credit management, reduces credit risk, and rationally allocates credit resources.
[0052] First, the process enters the multi-dimensional user credit data collection and fusion stage (S100). In this embodiment, multi-channel data connections are established through distributed data interfaces, and standardized interfaces are used for data transmission to ensure the stability and accuracy of data transmission. Static data related to user finances are obtained, and dynamic consumption behavior data such as user purchase records, return frequency, and consumption amount distribution are collected from e-commerce platforms. Social network data such as user friend interaction activity and social circle stability are collected from social media platforms. The collected data has various formats and quality issues, so it needs to be cleaned and normalized to map data with different characteristics to the same numerical range, eliminate the influence of units, and facilitate subsequent analysis. After cleaning and normalization, feature engineering techniques are used to extract key features from the multi-dimensional data, such as consumption frequency, average consumption amount, and diversity of consumption categories from consumption behavior data, and the number of friends and frequency of interaction from social network data. These features are combined and encoded to generate a vectorized feature set for user credit profiles, providing a data foundation for subsequent credit scoring modeling.
[0053] Then, the process moves to the risk-linked credit scoring modeling stage (S200). The vectorized feature set generated in S100 is used as input to the pre-trained risk assessment model. In this embodiment, the risk assessment model employs a deep neural network and collects a large amount of historical data. This historical data covers credit-related information of different users over a past period, divided into fixed time windows. Data within each time window includes the user's historical static data, scenario data, and user behavior time-series data for that period. The collected historical data is preprocessed to remove duplicate data, handle missing values and outliers, and is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used to train the model parameters, the validation set is used to evaluate the model performance during training and prevent overfitting, and the test set is used to finally evaluate the model's generalization ability. The deep neural network consists of an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the vectorized feature set generated by S100 to ensure complete reception of user credit feature data. The hidden layer contains multiple neurons, and an activation function is used to perform a non-linear transformation on the input data. The activation function f(x) = max(0,x) effectively alleviates the gradient vanishing problem and accelerates model convergence. When the input x of a neuron is less than 0, the output is 0; when x is greater than or equal to 0, the output is equal to x. In the hidden layer, neurons operate according to formula z... l =W l a l-1 +b l Perform a weighted summation calculation, where z l W is the weighted input of the l-th layer. l It is the weight matrix of the l-th layer, a l-1 It is the output of the previous layer, b l It is the bias vector of the l-th layer, and z l Input activation function f to obtain the output a of layer l. l =f(z) l The number of neurons in the output layer is set according to the number of risk level categories. For example, if the risk level is divided into three levels—low, medium, and high—then K=3. The output layer uses the softmax function for normalization. The formula for the softmax function is: Where z is the input vector of the output layer neuron, K is the number of output layer neurons, and σ(z) j Let represent the probability of the j-th neuron's output. The softmax function transforms the input of the output layer into a probability distribution, ensuring that the sum of the output probabilities of all neurons is 1. This yields the probability distribution of a user belonging to different risk levels. During model training, the cross-entropy loss function is used as the optimization objective. The formula for the cross-entropy loss function is: Where N is the number of training samples, K is the number of risk level categories, and yqj This represents the true label of the q-th sample belonging to the j-th risk level; 1 indicates it belongs to that level, and 0 indicates it doesn't. Let represent the probability that the model predicts the q-th sample belongs to the j-th risk level. In each training iteration, a batch of samples is randomly selected from the training set. For this batch of samples, the gradient of the loss function with respect to the model parameters weights W and bias b is calculated. The gradient calculations are as follows: The model parameters are updated based on the gradient, and the parameter update formula is as follows: Here, α is the learning rate, used to control the step size of parameter updates. During training, the learning rate and other hyperparameters (such as the number of hidden layer neurons and the number of layers) are dynamically adjusted by evaluating model performance on the validation set to prevent overfitting or underfitting. When the model's performance on the validation set no longer improves, training stops, resulting in a pre-trained risk assessment model. The feature set generated by S100 is input into the pre-trained risk assessment model. The data starts from the input layer, goes through layers of calculation and nonlinear transformation in the hidden layers, and finally reaches the output layer. The output layer uses the softmax function to obtain the probability distribution of users belonging to different risk levels, selects the risk level with the highest probability as the user's risk level, and maps this probability value to a preset risk range of five levels, A to E. For the calculation of the user risk fluctuation coefficient, based on the input user behavior time series data features, the standard deviation in time series analysis is used to measure the degree of fluctuation of user behavior. For example, the user's behavioral feature sequence in time period t is x1, x2, ..., x n The mean of this feature sequence is Then the risk volatility coefficient of this feature The risk volatility coefficient of a user is obtained by combining the risk volatility coefficients of multiple behavioral characteristics and averaging them according to the weights set for each behavioral characteristic in the risk assessment.
[0054] Subsequently, the system enters the dynamic matching phase of pre-approved credit limits and interest rates (S300). Based on the user risk level obtained in S200, the pre-approved credit rule engine is invoked to generate an initial credit limit range. The pre-approved credit rule engine is configured based on a large amount of historical data and business experience, mapping risk levels to credit limit ranges and setting a corresponding lower limit L for the initial credit limit range for each risk level. min and upper limit L max For example, for a user with a risk level of A, the lower limit of the initial credit limit range. Related to the user's historical annual income, upper limit Similarly, a certain percentage is set based on the user's historical annual income. The specific percentage is determined comprehensively based on factors such as the repayment ability and default probability of users with different risk levels in historical data. This aims to ensure that the credit limit setting both meets the user's reasonable needs and controls the financial institution's risk. The dynamic interest rate fluctuation range is determined based on the risk volatility coefficient and the market benchmark interest rate. When the market benchmark interest rate is r0 and the risk volatility coefficient is σ, a different interest rate adjustment coefficient k is set for each risk level. The higher the risk level, the larger the value of k. The dynamic interest rate lower limit r... min = r0 + k × σ, dynamic interest rate ceiling r max = r0 + 2k × σ, thus obtaining the dynamic interest rate fluctuation range (r min ,r max Meanwhile, to further control risk, when the risk volatility coefficient σ exceeds a preset threshold σ... th When a supplementary material application instruction is triggered, it is because an excessively high risk volatility coefficient may indicate that the user's credit status is unstable and more information is needed to accurately assess the risk. At this time, the user is required to submit supplementary materials such as income statements, asset documents, and third-party guarantee information. During the supplementary material application stage, the initial credit limit range is frozen and a 72-hour countdown begins. If the user fails to submit the materials within the time limit, the credit granting process is automatically closed. This is to avoid the credit granting process being extended indefinitely due to user delays, which would waste the resources of financial institutions. When the materials are approved, the risk score of S200 is updated based on the new data and S300 is re-executed to ensure that the determination of the credit limit and interest rate is more accurate and reasonable.
[0055] Next, the process moves to the credit limit fitting and dynamic adjustment generation stage (S400). This involves constructing a multi-objective optimization function to determine the optimal dynamic credit limit. The multi-objective optimization function sets a risk coverage objective R and a profit maximization objective P. The risk coverage objective R measures the degree to which the credit limit covers risk, and the formula is as follows: Among them, w u It is the weight of the u-th risk level, g u This formula represents the risk probability corresponding to the u-th risk level, where U is the number of risk levels. It reflects the overall risk coverage by weighted averaging the risk probabilities of different risk levels, with weight w. u Based on the impact of different risk levels on overall risk in historical data, the profit maximization objective P is determined by considering factors such as loan interest income and funding costs. The formula is as follows: where l v This is the limit for the vth loan product, r v It is the interest rate of the vth loan product and r v ∈[r min ,r max ], c vLet V be the funding cost of the v-th loan product, and V be the number of loan products. This formula calculates the total profit of different loan products after considering interest rates and costs, in order to maximize the profit of financial institutions. The multi-objective optimization function is F = αR + βP, where α and β are weight coefficients used to balance the importance of each objective, and α + β = 1. The values of α and β are determined according to the business strategy and risk appetite of financial institutions. If financial institutions pay more attention to risk control, the value of α is appropriately increased; if they pursue profit more, the weight of β is increased. The multi-objective optimization function F is solved. In the solution process, the initial quota interval L generated in S300 is considered. min ,L max and interest rate range [r] min ,r max By adjusting the combination of credit limit and interest rate, the optimal dynamic credit limit L is determined by finding the combination that maximizes the multi-objective optimization function F. op This paper introduces a real-time market interest rate factor M and a user behavior feedback factor B. The real-time market interest rate factor M reflects real-time changes in market interest rates and is calculated by collecting interest rate data from similar financial products in the market and performing a weighted average. The weights of the weighted average are determined based on factors such as the market share and influence of each financial product. The user behavior feedback factor B comprehensively considers the user's actual behavior after obtaining the pre-approved credit line. When the user repays on time, B = 1; otherwise, it is 0. Based on the user's behavioral data, the user's credit status is assessed in real time, and the assessment result is transformed into the user behavior feedback factor B. The final dynamic adjustment amount ΔL = L op ×(M+B) is a formula that enables dynamic adjustment of credit limits, making credit limits more closely match the user's actual credit status and market changes.
[0056] Finally, the closed-loop feedback and model iteration phase (S500) begins. After a user obtains a credit line and conducts credit transactions, the system continuously collects data on the user's actual credit performance, such as repayment records, credit line usage frequency, and overdue repayments. This data is collected and processed in real time, feeding back the actual credit performance data to the S500 risk assessment model and the S400 optimization algorithm. For the risk assessment model, the new credit performance data is merged with historical data, the training set, validation set, and test set are redefined, and the model is fine-tuned according to the previous training process, updating the model parameters to better adapt to the new data distribution and user behavior. For the S400 optimization algorithm, the weight coefficients α and β of the risk coverage objective R and the profit maximization objective P, as well as other relevant parameters, are adjusted based on new credit performance data to optimize the solution of the multi-objective optimization function. This further improves the accuracy of dynamic credit limits and adjustments. Through this online learning mechanism, model parameters are continuously updated, forming a closed-loop link of credit assessment, credit decision-making, dynamic adjustment, and feedback optimization. With the passage of time and data accumulation, the ability to assess users' creditworthiness can be continuously improved, credit limits and interest rates can be determined more accurately, credit risk can be effectively reduced, and the operational efficiency and service quality of financial institutions can be improved.
[0057] In summary, this embodiment fully demonstrates the operational process of a dynamic credit limit management method based on multi-dimensional user credit profiles. From multi-dimensional data collection and fusion to building credit profiles for risk assessment, and then to matching pre-approved credit limits with interest rates, dynamic adjustments, and closed-loop feedback optimization, each link is closely connected and collaborates with each other. By comprehensively considering multiple factors, it achieves accurate assessment of user credit status and dynamic management of credit limits, providing financial institutions with more scientific and efficient means in risk control and resource allocation.
[0058] Example 2
[0059] like Figure 2 As shown, this embodiment provides a dynamic credit limit management system based on multi-dimensional user credit profiles. It can accurately assess user credit, determine credit limits, and dynamically adjust them according to market changes and user behavior, effectively improving the risk management capabilities and service quality of financial institutions. The system mainly consists of a data acquisition and fusion module, a risk assessment module, a pre-credit decision module, a dynamic fitting module, and a closed-loop feedback module. The modules work together to form a complete closed-loop management system.
[0060] The data acquisition and fusion module deploys a distributed data crawler and API gateway, integrates internal and external data sources, and has a built-in feature engineering engine to output a vectorized feature set of user credit profiles.
[0061] The risk assessment module includes an offline training unit and an online prediction unit. The offline unit trains the risk assessment model using historical data, while the online unit calculates the user's risk score and risk volatility coefficient in real time and generates risk-driven decision signals.
[0062] The pre-credit decision module has a built-in credit rule engine that sets an initial credit limit range based on decision signals, supports manual intervention interfaces and automated material supplementation request generation, and outputs a credit limit and interest rate range pair with constraints.
[0063] The dynamic fitting module uses multi-objective optimization to perform secondary calibration on the pre-credit results, and combines real-time market data and user behavior data to output dynamic adjustment amounts and credit activation instructions for manual review.
[0064] The closed-loop feedback module collects user behavior data after credit authorization, constructs a feedback dataset, triggers a model retraining task, and synchronizes the updated model parameters to the risk assessment module and the dynamic fitting module.
[0065] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.
Claims
1. A dynamic credit limit management method based on multi-dimensional user credit profiles, characterized in that, The specific steps of this method are as follows: S100, Multi-dimensional User Credit Data Collection and Fusion: User credit feature data is acquired in real time through a distributed data interface, and feature engineering is used to clean and normalize the user credit feature data to generate a vectorized feature set for the user credit profile. S200, Risk-linked credit scoring modeling: Input the feature set of S100 into the pre-trained risk assessment model, and output the user's risk level and risk volatility coefficient; S300, dynamic matching of pre-approved credit limit and interest rate: Based on the risk level of S200, the credit rule engine is called to generate an initial credit limit range, and the dynamic interest rate fluctuation range is calculated in combination with the risk volatility coefficient; S400, Credit Limit Fitting and Dynamic Adjustment Generation: Multi-objective optimization is performed on the initial credit limit range and interest rate range of S300 to fit the dynamic credit limit, and the real-time market interest rate factor is introduced to generate the final dynamic adjustment amount. S500, closed-loop feedback and model iteration: The actual credit performance data is fed back to the risk assessment model of S200 and the optimization algorithm of S400. The model parameters are updated through an online learning mechanism, forming a closed-loop link of credit assessment - credit decision - dynamic adjustment - feedback optimization.
2. The dynamic credit limit management method based on multi-dimensional user credit profiles according to claim 1, characterized in that, The S100 user credit feature data includes static data, dynamic data, and scenario data, wherein: The static data includes: the user's age, gender, education, occupation, proof of income, proof of assets, and credit records; The dynamic data includes: real-time consumption behavior data, social network activity, mobile device usage behavior, and credit rating of social network contacts; The scenario data includes: the unemployment rate of the user's industry, the GDP growth rate of the macroeconomy, and the amount of regional credit adjustment.
3. The dynamic credit limit management method based on multi-dimensional user credit profiles according to claim 1, characterized in that, The risk assessment model is jointly trained using historical data and time-series user behavior data, dynamically adjusting the weight allocation of different dimensions to output the user's risk level and risk volatility coefficient. The training process is as follows: Training data preparation: Collect historical data and user behavior time series data. Divide the historical data into time windows. The data in each time window includes the user's historical static data and scene data within that time period. The user behavior time series data includes the user's historical dynamic data within that time period. Preprocess the historical data and divide it into training set, validation set and test set in a ratio of 7:2:
1. Risk assessment model construction: A deep neural network is used as the risk assessment model. This model consists of an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the vectorized feature set generated by S100, and it is used to receive user credit feature data. The hidden layers contain multiple neurons, and a non-linear transformation of the input data is achieved through an activation function, f(x) = max(0,x), where x is the input of the neuron. The number of neurons in the output layer is set according to the number of risk level categories; that is, the number of neurons in the output layer is the same as the number of risk level categories and is set to K. The output layer uses a softmax function for normalization. Where z is the input vector of the output layer neuron, K is the number of output layer neurons, and σ(z) j This represents the probability of the j-th neuron's output. This indicates that the input z of the j-th neuron in the output layer is... j Perform exponential operations with the natural constant e as the base. It applies to all K neurons in the output layer. The values are summed, and the probability distribution of users belonging to different risk levels is obtained through the softmax function. In the neural network, the output 'a' of each neuron is calculated from the input x, weight W, and bias b. For the l-th neuron, its z... l =W l a l-1 +b l a l =f(z) l ), where z l It is the weighted input of this layer, a l-1 It is the output of the previous layer, W l and b l These are the weight matrix and bias vector of the layer, respectively, and f is the activation function; Model training process: The cross-entropy loss function is used as the optimization objective. The cross-entropy loss function is... Where N is the number of training samples, L is the number of risk level categories, and y qj This represents the true label of the q-th sample belonging to the j-th risk level. Let represent the probability that the model predicts the q-th sample belongs to the j-th risk level. The model is trained using stochastic gradient descent. In each training iteration, samples are randomly selected from the training set, and the gradient of the loss function for each sample is calculated with respect to the model parameters weights W and bias b. The gradient is calculated as follows: The model parameters are updated based on the gradient. Where α is the learning rate, which controls the step size of parameter updates. By continuously adjusting the model parameters and evaluating the model performance on the validation set, a pre-trained risk assessment model is obtained.
4. The dynamic credit limit management method based on multi-dimensional user credit profiles according to claim 3, characterized in that, The risk level predicted by the model The specific calculation process is as follows: The feature set of the training samples is input into the input layer of the deep neural network. The input layer passes the data to the first hidden layer, and the hidden layer calculates the result based on z. l =W l a l-1 +b l and a l =f(z) l The input data is weighted, summed, and processed by an activation function, and then passed layer by layer until it reaches the output layer. At the output layer, the input vector z is obtained through weighted summation and then processed by the softmax function. The probability of the sample belonging to each risk level is calculated, and the vector composed of these probabilities is the model's prediction output, where the value of the j-th element is... The feature set generated by S100 is input into the pre-trained risk assessment model. The output layer obtains the probability distribution of the user belonging to different risk levels through the softmax function, selects the risk level with the highest probability as the user's risk level, and maps this probability value to the preset risk range of five levels from A to E. For the calculation of user risk volatility coefficient, based on the characteristics of the input user behavior time series data, the standard deviation in time series analysis is used to measure the degree of volatility of user behavior. That is, the user's behavioral characteristic sequence within this time period is x1, x2, ..., x n The mean of this feature sequence is Then the risk volatility coefficient of this feature This serves as the user's ultimate risk volatility coefficient.
5. The dynamic credit limit management method based on multi-dimensional user credit profiles according to claim 1, characterized in that, The S300 calls the pre-authorized credit rule engine to generate the initial credit limit range. The specific process is as follows: Risk Level to Credit Limit Range Mapping: The risk levels obtained from S200 are mapped to the credit limit ranges in the pre-approved credit rule engine. Based on historical data, the pre-approved credit rule engine sets a corresponding initial lower limit L for each risk level. min and upper limit L max Specifically: For users with a risk level of A, the lower limit of the initial credit limit range Up to 80% of the user's historical annual income. 120% of the user's historical annual revenue; For users with a risk level of B, the lower limit 60% of the user's historical annual income, with a maximum limit. 100% of the user's historical annual revenue; For users with a risk level of C, the lower limit 40% of the user's historical annual income, with a maximum of This represents 80% of the user's historical annual revenue. For users with risk level D, the lower limit 20% of the user's historical annual income, with a maximum of 60% of the user's historical annual revenue; For users with a risk level of E, the lower limit 10% of the user's historical annual income, with a maximum of This is 30% of the user's historical annual revenue.
6. The dynamic credit limit management method based on multi-dimensional user credit profiles according to claim 1, characterized in that, The S300 dynamic interest rate fluctuation range is determined based on the risk volatility coefficient and the market benchmark interest rate. When the market benchmark interest rate is r0 and the risk volatility coefficient is σ, the interest rate adjustment coefficient is k for each risk level. For users with risk level A, k... A =1.0, for users with risk level B, k B =1.2, for users with risk level C, k C =1.4, for users with risk level D, k D =1.6, for users with risk level E, k E =1.8; For the dynamic interest rate lower bound r min = r0 + k × σ, dynamic interest rate ceiling r max =r0 + 2k × σ, thus obtaining the dynamic interest rate fluctuation range (r min r max ); At the same time, when the risk volatility coefficient σ is greater than the preset threshold σ th The system triggers a supplementary material application instruction, requiring users to submit proof of income, asset collateral documents, and third-party guarantee information. During the supplementary material application stage, the initial credit limit range is frozen and a 72-hour countdown begins. If the user fails to submit the materials within the time limit, the credit granting process is automatically closed. When the materials are approved, the risk score of S200 is updated based on the new data, and S300 is re-executed.
7. The dynamic credit limit management method based on multi-dimensional user credit profiles according to claim 1, characterized in that, The S400 performs multi-objective optimization by constructing a multi-objective optimization function. This function sets a risk coverage objective R and a profit maximization objective P. The risk coverage objective R measures the degree to which the credit line covers risk. Where w u It is the weight of the u-th risk level, g u This represents the risk probability corresponding to the u-th risk level, where U is the number of risk levels. The profit maximization objective P is used to comprehensively consider factors such as loan interest income and funding costs. Among them l v This is the limit for the vth loan product, r v It is the interest rate of the vth loan product and r v ∈[r min ,r max ], c v Let be the cost of funds for the v-th loan product, and V be the number of loan products. Then the multi-objective optimization function is F = αR + βP, where α and β are weight coefficients used to balance the importance of each objective, and α + β = 1. The multi-objective optimization function F is solved, and during the solution process, the initial quota interval [L] generated in S300 is used. min ,L max and interest rate range [r] min ,r max By adjusting the combination of credit limit and interest rate, the multi-objective optimization function F is maximized to determine the optimal dynamic credit limit L. op ; A real-time market interest rate factor and a user behavior feedback factor are introduced. The real-time market interest rate factor M reflects the real-time changes in market interest rates and is obtained by collecting interest rate data of similar financial products in the market and performing a weighted average calculation. The user behavior feedback factor B integrates the user's actual behavior after obtaining the pre-approved credit limit; that is, B = 1 when the user repays on time and 0 otherwise. Based on the user's behavior data, the user's credit status is assessed in real time, and the assessment result is transformed into the user behavior feedback factor B. The final dynamic adjustment amount ΔL = L op ×(M+B).
8. A dynamic credit limit management system based on multi-dimensional user credit profiles, applicable to the dynamic credit limit management method based on multi-dimensional user credit profiles as described in any one of claims 1-7, characterized in that, The system consists of: a data acquisition and fusion module, a risk assessment module, a pre-credit decision module, a dynamic fitting module, and a closed-loop feedback module. The data acquisition and fusion module deploys a distributed data crawler and API gateway, integrates internal and external data sources, and has a built-in feature engineering engine to output a vectorized feature set of user credit profiles. The risk assessment module includes an offline training unit and an online prediction unit. The offline unit trains the risk assessment model using historical data, while the online unit calculates the user's risk score and risk volatility coefficient in real time and generates risk-driven decision signals. The pre-credit decision module has a built-in credit rule engine that sets an initial credit limit range based on decision signals, supports manual intervention interfaces and automated material supplementation request generation, and outputs a credit limit and interest rate range pair with constraints. The dynamic fitting module uses multi-objective optimization to perform secondary calibration on the pre-credit results, and combines real-time market data and user behavior data to output dynamic adjustment amounts and credit activation instructions for manual review. The closed-loop feedback module collects user behavior data after credit authorization, constructs a feedback dataset, triggers a model retraining task, and synchronizes the updated model parameters to the risk assessment module and the dynamic fitting module.